LEARNING (Back-propagation)
| Collection | International Encyclopedia of Systems and Cybernetics |
|---|---|
| Year | 2004 |
| Vol. (num.) | 2(1) |
| ID | ◀ 1876 ▶ |
| Object type | Discipline oriented, General information |
Learning seems to be basically a recurrent two stages process in neural networks, a model of which could be the back-propagation algorithm in neural computer networks.
According to G. HINTON “The most serious objection to back-propagation as a model of real learning is that it requires a teacher to supply the desired output for each training example. In contrast, people learn most things without the help of a teacher. Nobody presents us with a detailed description of the internal representation of the world that we must learn to extract from our sensory input.” (1992, p.108).
It could be argued that living systems have such a “teacher”, in the guise of their genetic structure, which allows them to start from basic physiological standards, genetically embedded, that can be “educated”. The natural respiratory rhythm or the sense of equilibrium seem to be good examples. Learning however, to ride a bicycle is basically a self-correcting back-propagating process.
The general sequential learning process should thus be:
- put to use or establish a basic standard
- calculate the output deviation
- correct the inputs in order to obtain a better correspondance with the standards
At a more complex level, a process of correcting the standards themselves may appear, specially behavioral, conceptual or ethical ones, through ethological or cultural interaction.